MLIR News, 27th edition (2/19/2020)

[mehdi_amini]

See the previous published edition.

Welcome to the twenty-seventh issue of the MLIR (bi)Weekly, a newsletter covering developments in MLIR, and related projects in the ecosystem. MLIR (bi)Weekly is brought to you by a collective effort of contributors, we welcome your contributions!

Highlights

• The first end-to-end milestone of the sparse compiler was presented last week: see “Recent Talks” at the bottom and the Discourse thread: MLIR Support for Sparse Tensors

MLIR Core

Infrastructure

• Dialect interfaces can now be registered independently from the the dialect before loading. The dialect does not need to know about the interface, enabling simple extensions to existing dialects.
• Translation to LLVM IR is now based on dialect interfaces and is therefore extensible to any dialects, including out-of-tree ones.
• Initial alias-analysis framework has landed.
• The C API can now walk the IR up the tree
• A floating point fused multiply-add operation was added to the standard dialect.
• Debug Actions infrastructure in MLIR is landing.
• Debug Counters are now supported in MLIR using the debug action framework.
• Attribute/Type::getChecked and verifyConstructionInvariants are being refactored to be cleaner and more efficient.

Table-driven Infrastructure

• Declarative operation format specification now supports a general ref directive, replacing type_ref, that allows for referencing most components of an operation within a custom directive.

Optimizations and Code Generation

• Affine loop fusion now supports producer-consumer loop fusion with multiple producer stores. This new feature extends loop fusion applicability to more generic cases: https://reviews.llvm.org/D92876
• Vector contract and outerproduct received the addition of a “kind” to specify the “join” operation performed in the reduction.
• A new linal.tiled_loop operation was added, see [RFC] Add Linalg TileOp for more info.
• Multiple lowering were added to map the TOSA dialect to Linalg (here for example).

CPU codegen

• A pass to provide expansions of math operation into fast polynomial approximation was added, this is based on Eigen and Cephes.

SPIR-V

• More types and ops were defined to support graphics use cases.
• More patterns were added to convert vector ops to SPIR-V ops.
• SPIR-V serializer was split into multiple source files, in preparation for later refactoring to have better internal structures.

Other

• The support for the !llvm prefix for built-in type is removed. As a reminder the LLVM dialect is now using directly the builtin types when possible.
• The split of the standard dialect continues with the extraction of the math and memref dialects.
• The C++ runtime for the MLIR JIT/ExecutionEngine gets additions:
  • More helpers to manipulate/populate MemRef in C++
• Python bindings were added for AffineMap

In the Ecosystem

Flang, the LLVM Fortran Compiler

• Support for -fopenmp and -fopenacc was added into the flang driver.
• fir-opt tool was added, it will be used for passes and roundtrip testing.

TensorFlow / MLIR-HLO

• XLA GPU backend: All emitters are migrated to take LMHLO as input (individually, still need to glue them together):
  • Fused Slice
  • CollectivePermute
  • FFT
  • TriangularSolve
  • Fused Scatter
  • Bitcast
• Kernel Generator:
  • Multiple performance tweaks that fix the broadcasting performance.
  • Identified a missing host side optimization that Eigen performs: In broadcasting cases, Eigen computes a minimal broadcasted shape to reduce the number of dimensions of the tensor that is being computed. We are adding support for this, as well.
  • We have launched a few more unary kernels and our first launched binary kernel is faring well. We will start landing the other ones next.

CIRCT : Circuit IR Compilers and Tools aka ‘MLIR for hardware’

• A new comb dialect is created and all the RTL combinational ops moves into it.
• During ODMs John Demme presented Endianness rationale and Rajit Manohar talked about Asynchronous design.

Recent Talks

• 2021-02-11: Polygeist: An affine C frontend for MLIR ; slides - recording
• 2021-02-18 Compiler Support for Sparse Tensor Computations: A Status Update ; slides - recording

Recent Publications

• A High-Performance Sparse Tensor Algebra Compiler in Multi-Level IR ; this is sparse extension to COMET (see previous presentation: “Domain Specific Compilation for Computational Chemistry” ; slides - recording)